Method, System, Storage Medium and Electronic Device for Extracting Water Networks from Polarimetric SAR Images

Abstract

The present invention discloses a method, system, storage medium and electronic device for extracting water networks from polarimetric SAR images, including the following steps: preprocessing the SAR image to obtain backscatter images of different polarization modes and different polarization component images; dividing the acquired images into a training image set and a validation image set, and manually annotating the training image set to obtain a correct water network label set; constructing a deep encoding network and a decoding network, training and testing the model, and finally, according to requirements, the extraction result. By making full use of the polarimetric scattering information in the SAR image and the feature extraction ability of deep learning, the present invention builds a water network extraction model with strong generalization and robustness, that is, integrating the SAR polarization information into the deep learning method to meet the high-precision water body extraction in SAR images with high scene complexity, rich ground object types and wide swaths, thereby constructing a more accurate and complete water network, greatly improving the accuracy and precision of water network extraction.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and particularly to a method, a system, a storage medium and an electronic device for extracting water networks from polarimetric SAR images. Background Art

[0002] At present, due to its advantages of all-weather and all-day operation, Synthetic Aperture Radar (SAR) is widely used in fields such as agricultural management and disaster monitoring. Since the discrimination features exhibited by water bodies in SAR remote sensing images are relatively significant, and the research on water network detection in SAR images is of great significance in surface water monitoring, change detection of rivers and lakes, flood monitoring, comprehensive surveys and resource environment planning, its research and application have received extensive attention.

[0003] How to quickly and accurately extract water body information from synthetic aperture radar images and form a water network can effectively provide a basis for subsequent research in fields such as surface water monitoring, change detection of rivers and lakes, and flood monitoring.

[0004] The applicant's research on the existing technology found that the current various SAR image water body extraction methods can be mainly divided into extraction methods based on threshold segmentation, extraction methods based on active contour models, extraction methods based on Wishart distribution, and extraction methods based on deep learning.

[0005] The extraction method based on threshold segmentation determines whether the gray value of a pixel point in the SAR image meets the requirements of a preset threshold, so as to further determine whether the pixel point belongs to a water body or a background area. This method has the advantages of simple operation and high calculation efficiency. However, the selection of the threshold and the inherent speckle noise in the SAR image limit the use of this method.

[0006] The SAR image water body extraction method based on the Active Contour Mode (ACM) is mainly completed through two steps: rough segmentation and fine segmentation. First, the water body extraction result of rough segmentation is obtained, and then dynamic iteration is continuously performed according to the gradient information of the edge, and finally a fine water body extraction result is obtained. This method can also obtain a continuous, smooth and closed result in the case of high noise. However, its calculation is relatively complex, and the parameters need to be continuously adjusted and optimized.

[0007] The extraction method based on Wishart distribution classifies the SAR image by describing the covariance matrix of multivariate normal distribution samples. This method utilizes the information of the covariance matrix in the SAR image and can obtain a relatively accurate extraction result in areas with a small area. However, the spatial continuity of the result is poor.

[0008] The extraction method based on deep learning creates training samples and labels manually, designs a network model, trains the network model, and constructs a mapping model that can adapt to all pixels by adjusting the shared weights, and then extracts the water bodies in the input image. This method can effectively avoid the influence of speckle noise in SAR images, establish a more accurate mapping relationship, and thus obtain more accurate extraction results. However, this method does not make full use of the polarization information in SAR images, and a large number of sample sets need to be made in the early stage, resulting in a high labor cost.

[0009] It can be seen that the existing SAR image water body extraction technologies are affected by the inherent speckle noise in SAR images. It is difficult for traditional extraction methods based on threshold segmentation, Active Contour Mode (ACM), Wishart distribution, etc. to establish an accurate mapping relationship to extract water bodies, so a relatively complete water network cannot be constructed. Deep learning methods can effectively extract surface water bodies and thus construct a relatively complete water network. At present, most extraction methods based on deep learning mainly utilize the intensity information of SAR images and do not make full use of relevant polarization information, so there is room for further improvement in the accuracy and precision of water network extraction.

[0010] Therefore, how to integrate SAR polarization information into deep learning methods to meet the requirements of high-precision water body extraction in SAR images with high scene complexity, rich ground object types, and large swath widths, so as to construct a more accurate and complete water network is an urgent problem to be solved at present. Summary of the Invention

[0011] The purpose of the present invention is to provide a method, system, storage medium and electronic device for extracting water networks from polarimetric SAR images, which can make up for the defects of existing algorithms, solve the drawbacks of current water network extraction methods, and construct a more accurate and complete water network.

[0012] The technical solution adopted by the present invention is as follows:

[0013] A method for extracting water networks from SAR images based on a deep encoder-decoder network and polarization decomposition, the method comprising:

[0014] Step S101, preprocess the SAR image so that each pixel of the SAR image represents the true radar backscattering coefficient, and obtain backscattering images of different polarization modes;

[0015] The preprocessing includes orbit correction, thermal noise removal, radiometric calibration, Deburst, multi-look, filtering, and terrain correction;

[0016] Step S102, add an operation of generating a polarization matrix C2 after Deburst in step S101 to obtain a C2 matrix;

[0017] Step S103: Perform polarization decomposition on the C2 matrix based on a physical model to generate images of different polarization components. The C2 matrix can represent the polarization scattering information of each pixel in the image. By performing polarization decomposition on the C2 matrix of each pixel, different polarization scattering components of that point can be obtained, thereby obtaining images of different polarization scattering components.

[0018] Step S104: Divide the obtained backscattering image and polarization component images into a training image set and a validation image set. Then, manually annotate the training image set to obtain a correct water network label set.

[0019] Step S105: Construct a deep encoding network, including a feature extraction network module, a feature fusion network module, and a max pooling network module. Constructing a deep encoding network can effectively extract the features of the backscattering image and polarization component images at different resolution scales, and fuse the features of the backscattering image and polarization component images to obtain feature maps of different resolutions containing more information.

[0020] Step S106: Construct a deep decoding network, including a fusion module and a decoding network module, to obtain feature maps of different resolution fusions and water body position extraction information. The deep decoding network can effectively fuse feature maps of different resolutions to improve the model's detection ability for targets of different sizes, and at the same time output the water body position information predicted by the network.

[0021] Step S107: Model training. By inputting the training image set, validation image set, and label set, train the deep encoding and decoding fusion network model, and save the optimal network model. The constructed training image set and label set are input into the established network model for training, and the validation image set is input to verify the accuracy of the trained network model architecture, and save the network model parameters with the highest accuracy.

[0022] Step S108: Test the network model. Input the SAR image into the saved optimal network model to obtain the water network extraction result.

[0023] Step S109: Save or output the extraction result according to requirements.

[0024] In step S102, the preprocessing operation can generate the polarization matrix C2, that is, each pixel in the image has a 2*2 covariance matrix to represent, as shown in formula 1:

[0025]

[0026] where * represents conjugate transpose, c 11 、c 22 are both real numbers, c 12is a complex number; the C2 matrix is the basis for subsequent polarization decomposition based on the physical model.

[0027] In step S104, specifically: the acquired image is block-operated. A square of 1024*1024 is used to step in the horizontal and vertical directions in the image, and the step size is 512, obtaining overlapping block images; the blocked image is divided into a training image set and a validation image set according to a ratio of 8:2. The Labelme annotation software is used to frame and mark the water body position information in the training image set point by point along the water body edge to obtain a label set for storage.

[0028] The described deep encoding network structure is: the first convolutional layer → the first batch normalization layer → the first max pooling layer → the first activation function layer → the second convolutional layer → the second batch normalization layer → the second max pooling layer → the second activation function layer → the second data concatenation layer → the third matrix multiplication layer → the third batch normalization layer → the third activation function layer. Through the deep encoding network, feature maps with more information and different resolutions can be obtained.

[0029] The described deep decoding network structure is: the first matrix multiplication layer → the first batch normalization layer → the first activation function layer → the second matrix multiplication layer → the second batch normalization layer → the second activation function layer → the second data concatenation layer → the third matrix multiplication layer → the third batch normalization layer → the third activation function layer → the third data concatenation layer → the fourth upsampling layer → the fourth activation function layer. Through the deep decoding network, the extracted water body position information is finally obtained.

[0030] In step S107, the following formula is used as the model evaluation index;

[0031]

[0032]

[0033] In the formula, TP represents the number of pixel points that are actually water bodies and are predicted as water bodies by the network, FP represents the number of pixel points that are actually water bodies and are predicted as non-water bodies by the network, FN represents the number of pixel points that are actually non-water bodies and are predicted as water bodies by the network; Precision represents the accuracy rate; Recall represents the recall rate. By loading the trained model parameters, the accuracy of the images in the validation dataset is evaluated, the evaluation indexes of each model are statistically analyzed, and the model with the highest accuracy rate is saved as the optimized training model parameters.

[0034] In step S109, according to different requirements, the water network extraction results of the block image area can be output; or the extraction results of the block images can be superimposed to form the water network extraction results of the research area.

[0035] A water network extraction system for polarimetric SAR images based on a deep encoding and decoding network, comprising: a preprocessing unit configured to perform orbit correction, thermal noise removal, radiometric calibration, Deburst, generation of a polarization matrix C2, multi-look, filtering, and terrain correction operations on the SAR image to obtain backscatter images in different polarization modes and the polarization matrix C2;

[0036] A polarization decomposition unit configured to perform polarization decomposition on the polarization matrix C2 to obtain different polarization component images;

[0037] A deep encoding and decoding network unit configured to perform a deep encoding network and a deep decoding network, and perform feature extraction, feature fusion, etc. on the backscatter images in different polarization modes and different polarization component images to obtain an optimal network model and parameters;

[0038] A testing unit configured to extract water bodies from the input SAR image based on the optimal network model and parameters to obtain the water network extraction result of the SAR image based on the deep encoding and decoding network and polarization decomposition.

[0039] The program is any one of the water network extraction methods for polarimetric SAR images based on a deep encoding and decoding network.

[0040] An electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements

[0041] The water network extraction method for polarimetric SAR images based on a deep encoding and decoding network.

[0042] By fully utilizing the polarization scattering information in the SAR image and the powerful feature extraction ability of deep learning, the present invention constructs a water network extraction model with strong generalization and robustness, that is, integrating the SAR polarization information into the deep learning method to meet the high-precision water body extraction in SAR images with high scene complexity, rich ground object types, and large swath width, thereby constructing a more accurate and complete water network, further improving the accuracy and precision of water network extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0044] Figure 1 It is a schematic flow chart of the present invention;

[0045] Figure 2 The SAR image provided in the embodiment of the present invention;

[0046] Figure 3 The water body extraction result provided in the embodiment of the present invention;

[0047] Figure 4 The SAR image provided in the embodiment of the present invention;

[0048] Figure 5 The water network extraction result provided in the embodiment of the present invention. Detailed implementation manners

[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] As Figure 1 、 2 and shown in 3, the present invention includes the following steps:

[0051] Step S101: Preprocess the SAR image to make each pixel of the SAR image represent the true radar backscattering coefficient, and obtain the backscattering images of different polarization modes;

[0052] The preprocessing includes orbit correction, thermal noise removal, radiometric calibration, Deburst, multi-look, filtering, and terrain correction;

[0053] Step S102: Add an operation to generate the polarization matrix C2 after Deburst in step S101 to obtain the C2 matrix; in step S102, the preprocessing operation can generate the polarization matrix C2, that is, each pixel in the image has a 2*2 covariance matrix to represent, as shown in formula 1:

[0054]

[0055] Among them, * represents conjugate transpose, c 11 、c 22 are both real numbers, and c 12 is a complex number; the C2 matrix is the basis for subsequent polarization decomposition based on the physical model.

[0056] Step S103: Perform polarization decomposition on the C2 matrix based on a physical model to generate images of different polarization components. The C2 matrix can represent the polarization scattering information of each pixel point in the image. By performing polarization decomposition on the C2 matrix of each pixel point, different polarization scattering components of that point can be obtained, thereby obtaining images of different polarization scattering components.

[0057] Step S104: Divide the acquired backscattering image and polarization component images into a training image set and a validation image set, and then manually annotate the training image set to obtain a correct water network label set. Specifically in Step S104: Perform a blocking operation on the acquired images. Use a 1024*1024 square to step in the horizontal and vertical directions in the image, with a step size of 512, to obtain overlapping block images. Divide the blocked images into a training image set and a validation image set in a ratio of 8:2. Use the Labelme annotation software to frame and mark the water body position information in the training image set point by point along the water body edge to obtain a label set for storage.

[0058] The described deep encoding network structure is: First Convolutional Layer → First Batch Normalization Layer → First Max Pooling Layer → First Activation Function Layer → Second Convolutional Layer → Second Batch Normalization Layer → Second Max Pooling Layer → Second Activation Function Layer → Second Data Concatenation Layer → Third Matrix Multiplication Layer → Third Batch Normalization Layer → Third Activation Function Layer. Through the deep encoding network, feature maps with more information and different resolutions can be obtained.

[0059] The described deep decoding network structure is: First Matrix Multiplication Layer → First Batch Normalization Layer → First Activation Function Layer → Second Matrix Multiplication Layer → Second Batch Normalization Layer → Second Activation Function Layer → Second Data Concatenation Layer → Third Matrix Multiplication Layer → Third Batch Normalization Layer → Third Activation Function Layer → Third Data Concatenation Layer → Fourth Upsampling Layer → Fourth Activation Function Layer. Through the deep decoding network, the extracted water body position information is finally obtained.

[0060] Step S105: Construct a deep encoding network, including a feature extraction network module, a feature fusion network module, and a max pooling network module. Constructing the deep encoding network can effectively extract the features of the backscattering image and polarization component images at different resolution scales, and fuse the backscattering image features and polarization component image features to obtain different resolution feature maps with more information.

[0061] Step S106: Construct a deep decoding network, including a fusion module and a decoding network module, to obtain feature maps with different resolution fusions and water body position extraction information. The deep decoding network can effectively fuse feature maps with different resolutions to improve the model's detection ability for targets of different sizes, and at the same time output the water body position information predicted by the network.

[0062] Step S107, model training: Input the training image set, validation image set, and label set to train the deep encoding-decoding fusion network model, and save the optimal network model; The constructed training image set and label set are input into the established network model for training, and the validation image set is input to verify the accuracy of the trained network model architecture, and save the parameters of the network model with the highest accuracy; Specifically, the accuracy and recall rate can be calculated respectively by comparing the extraction results obtained by the algorithm with the manually annotated labels; In step S107, the following formula is used as the model evaluation index,

[0063]

[0064] In the formula, TP represents the number of pixel points that are truly water bodies and are predicted as water bodies by the network, FP represents the number of pixel points that are truly water bodies and are predicted as non-water bodies by the network, and FN represents the number of pixel points that are truly non-water bodies and are predicted as water bodies by the network; Precision represents the accuracy; Recall represents the recall rate. By loading the trained model parameters, the accuracy of the images in the validation dataset is evaluated, the evaluation indexes of each model are statistically analyzed, and the model with the highest accuracy is saved as the optimal training model parameters.

[0065] Step S108, test the network model: Input the SAR image into the saved optimal network model, and the water network extraction result can be obtained;

[0066] Step S109, save or output the extraction result according to requirements.

[0067] In step S109, according to different requirements, the water network extraction result of the segmented image area can be output; or the extraction results of the segmented images can be superimposed to form the water network extraction result of the research area.

[0068] The present invention fully utilizes the polarization scattering information in the SAR image and the powerful feature extraction ability of deep learning to build a water network extraction model with strong generalization and robustness, that is, integrating the SAR polarization information into the deep learning method to meet the high-precision water body extraction in SAR images with high scene complexity, rich ground object types, and large swath width, so as to construct a more accurate and complete water network, further improving the accuracy and precision of water network extraction.

[0069] A polarization SAR image water network extraction system based on a deep encoding-decoding network, comprising: a preprocessing unit configured to perform orbit correction, thermal noise removal, radiometric calibration, Deburst, generate the polarization matrix C2, multi-look, filtering, and terrain correction operations on the SAR image to obtain the backscattering images of different polarization modes and the polarization matrix C2;

[0070] A polarization decomposition unit configured to perform polarization decomposition on the polarization matrix C2 to obtain images of different polarization components;

[0071] A deep encoding and decoding network unit configured to perform operations such as feature extraction and feature fusion on the backscattering images of different polarization modes and the images of different polarization components through a deep encoding network and a deep decoding network to obtain an optimal network model and parameters;

[0072] A testing unit configured to extract water bodies from the input SAR image based on the optimal network model and parameters to obtain the SAR image water network extraction result based on the deep encoding and decoding network and polarization decomposition.

[0073] A computer-readable storage medium storing a computer program thereon, and the program is the water network extraction method based on the deep encoding and decoding network and polarization SAR image as described above.

[0074] An electronic device includes: a memory, a processor, and a program stored in the memory and executable on the processor, and when the processor executes the program, it implements the water network extraction method based on the deep encoding and decoding network and polarization SAR image as described above.

[0075] Further, the following uses a specific example to illustrate this method. Specifically: Step S101: Preprocess the SAR image (orbital correction, radiometric calibration, Deburst, multi-look, filtering, terrain correction). Deburst is a specific step in SAR image processing that removes the invalid regions with no values in the image and merges the valid regions;

[0076] By performing orbital correction on the SAR image, a more accurate orbital file can be obtained to make subsequent processing more accurate; radiometric calibration can make the value of each pixel in the image represent the true radar backscattering value, and thus obtain an accurate backscattering image; Deburst can effectively remove the signal-free part in the Sentinel-1 IW SLC image; multi-look can effectively eliminate the speckle noise in the SAR image and greatly reduce the subsequent data volume; filtering can further eliminate the speckle noise in the SAR image to reduce the impact of speckle noise on subsequent processing; terrain correction, in addition to geocoding, also performs terrain radiometric correction processing on the SAR image to make the preprocessed image more conform to the actual ground objects. After preprocessing, backscattering images of two polarization modes are obtained.

[0077] Step S102: Preprocess the SAR image (orbital correction, radiometric calibration, Deburst, generate polarization matrix C2, multi-look, filtering, terrain correction). In Step S101 and Step S102, the filtering operation selects the Refined Lee filter with a window size of 7*7.

[0078] In step S102, the generation of the polarization matrix C2 is added to the preprocessing process of the SAR image. After this preprocessing step, each pixel in the image is represented by a 2×2 covariance matrix C2, as shown in Equation 1:

[0079]

[0080] where * represents the conjugate transpose, c 11 , c 22 are both real numbers, and c 12 is a complex number.

[0081] Step S103: Perform polarization decomposition on the C2 matrix based on a physical model to obtain the volume scattering polarization component m v and the remaining polarization components m s , and form images of different polarization components, and then extract the SAR polarization scattering information; specifically,

[0082] Convert the C2 matrix into a Stokes vector, as shown in Equation (2):

[0083]

[0084] In the formula, Re(c 12 ) represents the real part of c 12 , and Im(c 12 ) represents the imaginary part of c 12 .

[0085] Convert the Stokes vector into the sum of three polarization components, as shown in Equation (3):

[0086] S = m v s v + m s s p + n s n (6)

[0087] In the formula, n is the noise term and s n is a Stokes vector of random polarization, and the product of these two terms can be removed by the filtering operation in the preprocessing; m v is the volume scattering polarization component; m s is the remaining polarization component.

[0088] Therefore, only two terms need to be considered at this time, namely:

[0089]

[0090] In the formula, there are a total of 4 unknowns. The matrix G is introduced to solve them:

[0091]

[0092] In the formula, det is the function to calculate the determinant, and T is the matrix transpose.

[0093] This equation requires the following to hold:

[0094] ( s -m v s v ) T G( s -m v s v ) = 0 (9) Therefore, a univariate quadratic equation system about m v can be constructed, and then solved

[0095]

[0096] According to formula (6) and formula (10), the polarization components m v 、m s can be obtained.

[0097] Through the polarization decomposition of the SAR image polarization matrix C2, different polarization component maps are obtained. Step S104: Divide the data set into a training image set and a validation image set, and label the water bodies in the SAR images in the training image set to obtain a label set.

[0098] The backscattering image obtained in step S101 is segmented. A square of L*L is used to step in the horizontal and vertical directions in the image, and the step size is L / 2 to obtain overlapping segmented images. To effectively prevent overfitting, the segmented images are divided into a training image set and a validation data set according to a ratio of 8:2. The water bodies in the training image set are manually extracted, and a label set is formed and stored.

[0099] Step S105: Construct a depth encoding network

[0100] The deep encoding network consists of three stages: The first stage consists of a convolutional layer, a batch normalization layer, a max pooling layer, and an activation function layer: Among them, the convolutional kernel size of the convolutional layer is 3*3, and the number of kernels is 16; the stride of the max pooling layer is 2; the activation function layer selects the Relu function as the activation function. The second stage consists of a convolutional layer, a batch normalization layer, a max pooling layer, an activation function layer, and a data concatenation layer: Among them, the convolutional kernel size of the convolutional layer is 3*3, and the number of kernels is 32; the stride of the max pooling layer is 2; the activation function layer selects the Relu function as the activation function; the data concatenation layer fuses the feature maps to obtain a more informative feature map. The third stage takes the fused feature map as input and consists of a matrix multiplication layer, a batch normalization layer, and an activation function layer: Among them, the matrix multiplication layer scales the data in the way of variance scaling initialization; the activation function layer selects the Relu function as the activation function.

[0101] Step S106: Construct a deep decoding network

[0102] The deep decoding network consists of four stages: The first stage consists of a matrix multiplication layer, a batch normalization layer, and an activation function layer: Among them, the matrix multiplication layer scales the data in the way of variance scaling initialization; the activation function layer selects the Relu function as the activation function. The second and third stages both consist of a matrix multiplication layer, a batch normalization layer, an activation function layer, and a data concatenation layer: Among them, the matrix multiplication layer scales the data in the way of variance scaling initialization; the activation function layer selects the Relu function as the activation function; the data concatenation layer fuses feature maps of different resolutions. The fourth stage includes an upsampling layer and an activation function layer: The upsampling layer upsamples the finally obtained feature map to obtain a feature map with the same size as the input image; the activation function layer selects the Relu function (if the Relu value is equal to 0, it is a non-water body and is marked as 0; if the value is greater than 0, it is a water body and is marked as 1) to judge whether each pixel point in the feature map is a water body pixel to obtain the final extraction result.

[0103] Step S107: Train the network model based on the deep encoding-decoding network and polarization decomposition. By inputting the training image set, the validation image set, and the label set, train the deep encoding-decoding fusion network model and save the optimal network model; Input the constructed training image set and label set into the built network model for training, input the validation image set to verify the accuracy of the trained network model architecture, and save the network model parameters with the highest accuracy;

[0104] First, set the batch_size to 32, the maximum number of iterations to 200, and the learning rate to 0.001. Then, input the training dataset and the label set into the network model constructed in steps S105 to S106 to train the network, and test the network performance using the validation dataset every 50 generations and save the model parameters obtained from the current training.

[0105] Step S108: Test the network model based on the deep encoder-decoder network and polarization decomposition. Input the SAR image into the saved optimal network model to obtain the water network extraction result. Load the optimal model parameters saved in step S107, divide the SAR image into blocks according to step S102, and then input it into the trained network model. The output layer is the water body extraction result for each block.

[0106] Step S109: Save or output the extraction result according to requirements. According to different requirements, the water network extraction result of the divided image area can be output; or the extraction results of the divided images can be superimposed to form the water network extraction result of the research area.

[0107] Figure 1 This is the flowchart of the water network extraction for SAR images in the present invention. Figure 2 It is a local area cropped from the Sentinel-1A input image in Kaifeng City, Henan Province, and the image size is 1024 * 1024 pixels. Figure 3 This is for the present invention Figure 2 in the water network extraction result, where the black color represents water bodies. Figure 4 It is the water network extraction result of the Sentinel-1A image in Kaifeng City, Henan Province, where the blue color represents water bodies. From Figure 3 and Figure 4 it can be seen that the present invention has a good water body extraction result in complex scenes and can form a relatively complete water network. Moreover, due to the high extraction accuracy and high processing efficiency of the present invention, it also has good practical value.

[0108] It should be noted that the terms "including" and "having" in the specification and claims of this application and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0109] Note that the above is only a preferred embodiment of the present invention and the application of technical principles. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, re-adjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the specific embodiments described herein. Without departing from the concept of the present invention, more other effective embodiments can also be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. Method for extracting water networks from polarimetric SAR images, characterized in that, The method includes: Step S101: Preprocess the SAR image so that each pixel of the SAR image represents the true radar backscattering coefficient, and obtain backscattering images in different polarization modes; The preprocessing includes orbit correction, thermal noise removal, radiometric calibration, Deburst, multi-look, filtering, and terrain correction; Step S102: Add an operation to generate the polarization matrix C2 after Deburst in Step S101 to obtain the C2 matrix; Step S103: Perform polarization decomposition on the C2 matrix based on a physical model to generate different polarization component images; the C2 matrix can represent the polarization scattering information of each pixel point in the image. By performing polarization decomposition on the C2 matrix of each pixel point, different polarization scattering components of this point can be obtained, thereby obtaining different polarization scattering component images; Step S104: Divide the obtained backscattering image and polarization component image into a training image set and a validation image set, and then manually annotate the training image set to obtain a correct water network label set; Step S105: Construct a deep encoding network, including a feature extraction network module, a feature fusion network module, and a max pooling network module; constructing a deep encoding network can effectively extract the features of the backscattering image and polarization component image at different resolution scales, and fuse the features of the backscattering image and polarization component image to obtain different resolution feature maps containing more information; Step S106: Construct a deep decoding network, including a fusion module and a decoding network module, to obtain feature maps fused at different resolutions and water body position extraction information; the deep decoding network can effectively fuse feature maps at different resolutions to improve the model's detection ability for targets of different sizes, and at the same time output the water body position information predicted by the network; Step S107: Model training, by inputting the training image set, validation image set, and label set, train the deep encoding and decoding fusion network model, and save the optimal network model; input the constructed training image set and label set into the established network model for training, input the validation image set to verify the accuracy of the trained network model architecture, and save the network model parameters with the highest accuracy; Step S108: Test the network model, input the SAR image into the saved optimal network model, and the water network extraction result can be obtained; Step S109: Save or output the extraction result according to requirements.

2. The method for extracting water networks from polarimetric SAR images according to claim 1, characterized in that In Step S102, the preprocessing operation can generate the polarization matrix C2, that is, each pixel point in the image is represented by a 2*2 covariance matrix, as shown in Formula 1: where * denotes conjugate transpose, c 11 and c 22 are both real numbers, and c 12 is a complex number; the C2 matrix is the basis for subsequent polarization decomposition based on the physical model.

3. The method for extracting water networks from polarimetric SAR images according to claim 1, wherein Step S104 is specifically: perform a blocking operation on the obtained image, use a 1024*1024 square to step in the horizontal and vertical directions in the image, and the step size is 512 to obtain overlapping block images; divide the blocked image into a training image set and a validation image set according to a ratio of 8:2, and use the Labelme annotation software to select and mark the water body position information in the training image set point by point along the water body edge to obtain a label set for storage.

4. The method for extracting water networks from polarimetric SAR images according to claim 1, wherein The described deep encoding network structure is: First Convolutional Layer → First Batch Normalization Layer → First Max Pooling Layer → First Activation Function Layer → Second Convolutional Layer → Second Batch Normalization Layer → Second Max Pooling Layer → Second Activation Function Layer → Second Data Concatenation Layer → Third Matrix Multiplication Layer → Third Batch Normalization Layer → Third Activation Function Layer. Through the deep encoding network, feature maps with more information and different resolutions can be obtained.

5. The method for extracting water networks from polarimetric SAR images according to claim 1, wherein, The described deep decoding network structure is: First Matrix Multiplication Layer → First Batch Normalization Layer → First Activation Function Layer → Second Matrix Multiplication Layer → Second Batch Normalization Layer → Second Activation Function Layer → Second Data Concatenation Layer → Third Matrix Multiplication Layer → Third Batch Normalization Layer → Third Activation Function Layer → Third Data Concatenation Layer → Fourth Upsampling Layer → Fourth Activation Function Layer. Through the deep decoding network, the extracted water body position information is finally obtained.

6. The method for extracting water networks from polarimetric SAR images according to claim 1, wherein In step S107, the following formula is used as the model evaluation index; In the formula, TP represents the number of pixel points that are actually water bodies and are predicted as water bodies by the network, FP represents the number of pixel points that are actually water bodies and are predicted as non-water bodies by the network, and FN represents the number of pixel points that are actually non-water bodies and are predicted as water bodies by the network; Precision represents the accuracy rate; Recall represents the recall rate; By loading the trained model parameters, the accuracy of the images in the validation dataset is evaluated, the evaluation indexes of each model are statistically analyzed, and the model with the highest accuracy rate is saved as the optimized training model parameters.

7. The method for extracting water networks from polarimetric SAR images according to claim 1, characterized in that In step S109, according to different requirements, the water network extraction results of the segmented image regions can be output; or the extraction results of the segmented images can be superimposed to form the water network extraction results of the research area.

8. A water network extraction system for polarimetric SAR images, which is used to implement the water network extraction method for polarimetric SAR images according to any one of claims 1-7, characterized in that, Including: A preprocessing unit configured to perform orbit correction, thermal noise removal, radiometric calibration, Deburst, generate polarization matrix C2, multi-look, filtering, and terrain correction operations on the SAR image to obtain backscattering images in different polarization modes and polarization matrix C2; A polarization decomposition unit configured to perform polarization decomposition on the polarization matrix C2 to obtain different polarization component images; A deep encoding and decoding network unit configured to perform operations such as feature extraction and feature fusion on the backscattering images in different polarization modes and different polarization component images through the deep encoding network and the deep decoding network to obtain the optimal network model and parameters; A testing unit configured to extract water bodies from the input SAR image based on the optimal network model and parameters to obtain the water network extraction results of the SAR image based on the deep encoding and decoding network and polarization decomposition.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is the polarization SAR image water network extraction method as described in any one of claims 1-7.

10. An electronic device, characterized in that, Including: A memory, a processor, and a program stored in the memory and executable on the processor, and when the processor executes the program, it implements the polarization SAR image water network extraction method as described in any one of claims 1-7.

Images